Researchers Use Soft Prompts to Adapt AI Models to Different Cultural Values

Researchers have developed a novel way to make large language models behave in culturally appropriate ways by encoding values into specialized, swappable text prompts rather than retraining the underlying models. The approach keeps model weights frozen and requires no labeled preference data, making it practical for deployment.

The Challenge of Cultural Alignment#

Standard methods for aligning language models—like supervised fine-tuning or reinforcement learning—rely on updating model weights and need labeled training data. Yet deployed models often produce culturally homogeneous responses that don't reflect the diverse values across different societies. Aligning models to specific cultural frameworks, such as Hofstede's Value Survey Module (VSM13), poses a unique problem: alignment signals come only as aggregated scores from entire surveys, with no per-token guidance and no conventional preference data to work from.

A Lightweight Solution#

Rather than modify the model itself, the team encodes cultural behavior into short, learnable soft prompts—essentially prepended text sequences that guide model outputs. These prompts are optimized using Differential Evolution, a gradient-free search method that iteratively refines candidate solutions. At inference time, the system selects and inserts the appropriate cultural-specific prompt before generating responses, effectively adapting the model's behavior for different cultures.

Results Across Multiple Tests#

Testing spanned four countries and four instruction-tuned models. The researchers found that DE-optimized prompts generally reduced divergence from VSM13 cultural reference profiles. The approach also improved alignment with the World Values Survey (WVS)—an independent evaluation framework not used during optimization—suggesting the method captures genuine cultural adaptation rather than overfitting. In blinded pairwise comparisons judged by multiple language models, the culturally adapted outputs were preferred by majority vote.

Why It Matters#

This deployment-friendly approach sidesteps the need to retrain expensive models while making them more culturally responsive. By keeping weights static and relying only on lightweight prompt tuning, organizations can quickly adapt a single model for many different cultural contexts.